[Bug]: Closing one client from a ClientFactory closes the shared httpx client for every other client

已關閉
  • 哪裡不對？
#1,349 0 則留言 0 個 reaction 已指派 2 人 在 GitHub 檢視

維護者通常 2 天內回覆

@rohityan 已經在處理了。

開始於 2026年10月16日。

  • #1358 來自 @Iwaniukooo11 —— 已合併

評估

這個 Issue 還沒有評估資料。

描述

What happened?

Since #947, ClientFactory creates a single httpx.AsyncClient in __init__ (src/a2a/client/client_factory.py) and passes it to every transport it creates, and to the card resolver in create_from_url:

httpx_client = config.httpx_client or httpx.AsyncClient()
...
self._httpx_client = httpx_client

Each HTTP transport closes that client in close() (JsonRpcTransport.close, RestTransport.close and the v0.3 compat transports):

async def close(self) -> None:
    """Closes the httpx client."""
    await self.httpx_client.aclose()

So closing one client, explicitly or by leaving async with client:, closes the HTTP client shared by every other client from the same factory, and by the factory itself. Before #947 each transport got its own httpx.AsyncClient() unless one was passed in ClientConfig, so this did not happen.

Reproduction:

factory = ClientFactory()
client_a = factory.create(card_a)
client_b = factory.create(card_b)

async with client_a:
    ...

await client_b.get_task(GetTaskRequest(id='task-1'))
# RuntimeError: Cannot send a request, as the client has been closed.

await factory.create_from_url('https://agent-c.example')
# RuntimeError: Cannot send a request, as the client has been closed.

The error is a raw RuntimeError from httpx, not an A2AClientError, so it is also not caught by code handling SDK errors. The create_client docstring suggests using ClientFactory directly "for reusing a factory across multiple agents", which is the case that breaks.

Expected: closing a client releases only its own resources. Other clients from the same factory, and the factory itself, keep working.

I see two possible directions, depending on whether sharing the connection pool across clients was a goal of #947:

Option A: one HTTP client per created client (as before #947). When ClientConfig.httpx_client is not set, each create() builds its own httpx.AsyncClient (with the same default A2A-Version header), and create_from_url resolves the card with its own short-lived client. Small change, no new API, same behavior as before #947.

Option B: the factory owns the shared client. Transports stop closing an HTTP client they did not create, and ClientFactory gets close() and async with support to release the shared client. Keeps a single connection pool, but adds API surface and makes the factory responsible for its lifetime.

In both cases, a client passed through ClientConfig.httpx_client would be left to the caller to close. Today the SDK closes it too, which was already the case before #947, so I'm happy to keep that out of scope if you prefer.

Happy to open a PR with either option, or with a different approach if you have one in mind. Just let me know which direction you prefer.

Code of Conduct
  • I agree to follow this project's Code of Conduct
主要語言
Python
星號
2.2k
分支
532
平均合併
3 天 6 小時
30 天內合併 PR
42

環境準備

從這裡開始

  1. 先讀完整個 Issue，再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫，在一個分支上完成修改。
  4. 送出 Pull Request，並在描述裡引用這個 Issue 編號。

a2aproject/a2a-python 的其他 Issue

查看 a2aproject/a2a-python 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。